Soniox

What Soniox does

Soniox is a speech AI company founded by Klemen Simonic (CEO) and Ambroz Bizjak (Chief Architect). Its API does real-time (streaming) and asynchronous (file) transcription, speech translation and text-to-speech, with speaker diarization, language identification, endpoint detection and custom context. Its docs list 60 supported languages, which it markets as "60+". STT v6 (models stt-rt-v6 and stt-async-v6, released 3 Nov 2026) adds "agent text input": a voice agent streams the words it is speaking to Soniox on the same connection as the caller's audio, so the model knows what the caller is answering and can filter leftover echo of the agent's own voice out of the transcript. Soniox's docs say this works alongside normal echo cancellation and doesn't replace it. The v5 models will be retired on 15 Dec 2026 and requests will route to v6 automatically.

Good for

• Cheap real-time transcription for voice agents and phone bots
• Multilingual calls where people switch languages mid-sentence
• Live speech translation and meeting transcription on a budget

Honest take

No red flags in our notes so far

• Very cheap: about $0.12 an hour for real-time transcription, with diarization and translation included. Worth a test if speech-to-text is a big part of your voice-bot bill. We haven't tested it.
• STT v6's "agent text input" is useful for phone bots: the model knows what your agent just asked, so short answers like names, dishes and numbers should be recognised better, and leftover echo of the agent's voice is filtered out. To get the benefit, you have to send the agent's words in time with playback, about 0.5 to 1 second ahead, using TTS timestamps.
• "World's first" is overstated. AssemblyAI's streaming model has taken the agent's text as context since June 2026, and Speechmatics' voice SDK could already ignore an assistant's enrolled voice. Soniox says to keep normal echo cancellation on. It has published no v6 accuracy numbers yet; its benchmark page still shows v5.
• Upgrade note: v5 is retired on 15 Dec 2026 and traffic moves to v6 automatically. v6 writes prices as spoken ("15 dollars and 20 cents", not "$15.20"), which could break code that parses amounts.

Our Hype or Fact? checks

• Partly factSoniox STT v6: the "world's first" speech model that knows what your voice agent says and filters out its voice? 3 Nov

Pricing

Pay as you go, billed in tokens (soniox.com/pricing, checked 3 Nov 2026). Speech-to-text works out at about $0.10 per hour for async (file) and $0.12 per hour for real-time (streaming): audio input $1.50 (async) or $2.00 (real-time) per 1M tokens, text input and output $3.50 or $4.00 per 1M tokens (1 hour of audio is about 30,000 audio tokens). Diarization, language identification and translation are included. Agent text input is billed as input text tokens. Text-to-speech is about $0.70 per hour of generated speech ($4.00 per 1M input text tokens, $21.50 per 1M output audio tokens). Enterprise pricing on request.

Prices as listed by the maker when we last checked (2026-11-03); usually in US dollars. Check the website for current prices.

Latest update

• 3 November 2026: Released Soniox STT v6 (stt-rt-v6, stt-async-v6). New agent text input lets a voice agent stream what it is saying to the model, so it understands replies better and filters the agent's own voice out of the transcript. Soniox also claims significantly higher accuracy in 60+ languages and better diarization, language ID, endpointing and translation, but hasn't published v6 numbers. Our check rates the "world's first" claim as partly fact. v5 is retired on 15 Dec 2026.

Is Soniox worth it? Quick answers